CS Seminar: Large Scale EDA Optimisation

Advertisements for Seminars taking place on campus.
Forum rules
This forum is now unmoderated.

CS Seminar: Large Scale EDA Optimisation

Postby J. Borg » Tue Nov 08, 2016 12:35 pm

COMPUTING SEMINAR
WEDNESDAY 9TH NOVEMBER, 3:00PM, CR08
(Turing Lab, Colin Reeves Building)

Tea, Coffee and Biscuits available before, during and after the seminar
ALL WELCOME

"Large Scale EDA Optimisation by Aggregating Compressive Covariance Matrices"
Dr. Ata Kaban, School of Computer Science - University of Birmingham

ABSTRACT
Estimation of distribution algorithms (EDA) are a branch of population
based heuristic black-box optimisers that search for the global
optimum of a function by estimating and sampling a probabilistic model
of selected search points. While these methods have advantages in
principle over other evolutionary algorithms, their use in large scale
problems has been hindered by the curse of dimensionality in
multivariate model building. In this talk we consider large scale
continuous optimisation problems, and introduce random matrix theory
to EDA in order to devise scalable EDA-style methods. Model building
and sampling is done in low dimensional random
subspaces, and the aggregated covariance of the search distribution
enables exploration of the search space. We then show how we can
implicitly control the size of the aggregated covariance by changing
the distribution of the random matrices employed. Based on this we
develop adaptive schemes that balance exploration vs exploitation
during the search. Experimental tests confirm the effectiveness of our
approach, and we achieve state of the art performance in
1000-dimensional multi-modal benchmark problems.
J. Borg
 
Posts: 30
Joined: Fri Oct 23, 2009 3:58 pm

Return to Seminars

Who is online

Users browsing this forum: No registered users and 2 guests

cron